[rob-p]

Release 2.11.0

This merges develop into master for 2.11.0. The full notes are in docs/release-notes-2.11.0.md. The index format is unchanged.

Contents:

• perf(map): hold reference sequences at two bits per base while mapping #1265: two-bit reference sequences for selective alignment (Benjamin Demaille, with maintainer edits). Peak memory falls by 36–45% on GENCODE and 29% with a GRCh38 genome decoy, at the same speed.
• feat(io): read FASTQ and reference FASTA from streams #1269: quant and index read FIFOs and process substitution (Benjamin Demaille, with maintainer edits).
  • The truncation check moved from a separate pass over plain FASTQ into the reader. The reader comes from paraseq-temp 0.6.3-pre.1 via piscem-rs 0.12.3 (piscem-rs#19; offered upstream as Refuse a truncated FASTQ record; name the input in format errors noamteyssier/paraseq#102).
  • Format errors name their input, and plain FASTQ runs are about 10% faster.
• fix(map): do not reuse the previous fragment's hits for a mate shorter than k #1271, fix(quant): write a right orphan's SAM/BAM record on its own strand #1272: sketch mode no longer reuses the previous fragment's hits for a mate shorter than k; read-2 orphans are written to SAM/BAM on their own strand.
• Audit of 2.10.0 (Tracking: Ponytail audit findings (2.10.0) #1283), Benjamin Demaille: fix(index): a failed rebuild no longer destroys the existing index #1274–fix(rad): keep each orphan's own side and drop orphans beside a proper pair #1282, fix(cli): apply the -p cap to mapping threads, not only the rayon pool #1284–fix(align): warn when a -t FASTA sequence has another length than its reference #1300 and fix(map): keep a sketch right orphan's score on its SAM record #1302–perf(quant): add to the shared processed count once per batch #1308, several with maintainer edits; fix(quant): report the prior as the FLD source when no fragment length was observed #1301 is not included.
  • Wrong counts fixed, mostly in alignment mode (-a): orphan sides in RAD records (fix(rad): keep each orphan's own side and drop orphans beside a proper pair #1282, which also changes stranded reads-mode runs), single-end BAMs under -l A, SEQ * records under --errorModel, --hardFilter, outward library types.
  • Inputs that used to give a wrong result without an error are refused: unmatched decoy names, repeated reference names, k above 31, out-of-range flags, ambiguous quantmerge names, BAMs not grouped by read name.
  • A run that maps nothing exits non-zero, -a counts unmapped records in percent_mapped, and a rerun removes the old quant.sf first.
• docs: explain what inferential replicates do not measure #1266: inferential replicates docs: why salmon offers bootstrap and Gibbs, and what replicates do not measure.
• Fix: --online --sketch --posBias no longer panics.
• Dependencies: chore(deps): bump the cargo-patch group with 8 updates #1268 (Cargo patch updates), build(deps): bump rust from 1.98-bookworm to 1.99-bookworm #1262 (Docker Rust 1.99), chore(deps): bump the website-minor group in /website with 2 updates #1267 and chore(deps): bump smol-toml from 1.8.0 to 1.9.1 in /website #1248 (website).
• On this branch: the release notes.

Results. Against 2.10.0 on x86 (muscle, GEUVADIS, GEUVADIS with the genome decoy, selective alignment and sketch, at 1, 8 and 16 threads), quant.sf is identical in 31 of 38 runs. The other 7 are stranded muscle with selective alignment, which #1282 changes: 14 more fragments map, and 2,228 of 4.14M reads (0.05%) move between transcripts. Timing and memory tables are in the notes.

Checks:

• cargo test --release --workspace --locked: 435 passed. Clippy (-D warnings) and fmt are clean.
• ./scripts/bump_and_publish.sh 2.11.0 --dry-run packages and verifies all ten crates at 2.11.0.

After merging: run ./scripts/bump_and_publish.sh 2.11.0 --publish on master, set the GitHub release notes from the notes file, and open the master → develop sync PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_016brRcptooQkxWr6TyWnwAE